From: Joel Fernandes <joel@joelfernandes.org>
To: "Uladzislau Rezki (Sony)" <urezki@gmail.com>
Cc: LKML <linux-kernel@vger.kernel.org>,
"Paul E . McKenney" <paulmck@kernel.org>,
RCU <rcu@vger.kernel.org>, Steven Rostedt <rostedt@goodmis.org>,
Oleksiy Avramchenko <oleksiy.avramchenko@sonymobile.com>
Subject: Re: [PATCH 1/1] rcu/tree: support kfree_bulk() interface in kfree_rcu()
Date: Wed, 15 Jan 2020 20:14:10 -0500 [thread overview]
Message-ID: <20200116011410.GC246464@google.com> (raw)
In-Reply-To: <20191231122241.5702-1-urezki@gmail.com>
On Tue, Dec 31, 2019 at 01:22:41PM +0100, Uladzislau Rezki (Sony) wrote:
> kfree_rcu() logic can be improved further by using kfree_bulk()
> interface along with "basic batching support" introduced earlier.
>
> The are at least two advantages of using "bulk" interface:
> - in case of large number of kfree_rcu() requests kfree_bulk()
> reduces the per-object overhead caused by calling kfree()
> per-object.
>
> - reduces the number of cache-misses due to "pointer chasing"
> between objects which can be far spread between each other.
>
> This approach defines a new kfree_rcu_bulk_data structure that
> stores pointers in an array with a specific size. Number of entries
> in that array depends on PAGE_SIZE making kfree_rcu_bulk_data
> structure to be exactly one page.
>
> Since it deals with "block-chain" technique there is an extra
> need in dynamic allocation when a new block is required. Memory
> is allocated with GFP_NOWAIT | __GFP_NOWARN flags, i.e. that
> allows to skip direct reclaim under low memory condition to
> prevent stalling and fails silently under high memory pressure.
>
> The "emergency path" gets maintained when a system is run out
> of memory. In that case objects are linked into regular list
> and that is it.
>
> In order to evaluate it, the "rcuperf" was run to analyze how
> much memory is consumed and what is kfree_bulk() throughput.
>
> Testing on the HiKey-960, arm64, 8xCPUs with below parameters:
>
> CONFIG_SLAB=y
> kfree_loops=200000 kfree_alloc_num=1000 kfree_rcu_test=1
>
> 102898760401 ns, loops: 200000, batches: 5822, memory footprint: 158MB
> 89947009882 ns, loops: 200000, batches: 6715, memory footprint: 115MB
>
> rcuperf shows approximately ~12% better throughput(Total time)
> in case of using "bulk" interface. The "drain logic" or its RCU
> callback does the work faster that leads to better throughput.
Tested-by: Joel Fernandes (Google) <joel@joelfernandes.org>
(Vlad is going to post a v2 which fixes a debugobjects bug but that should
not have any impact on testing).
thanks,
- Joel
>
> Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
> ---
> kernel/rcu/tree.c | 154 ++++++++++++++++++++++++++++++++++++++--------
> 1 file changed, 130 insertions(+), 24 deletions(-)
>
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index 48fba2257748..4ee5c737558b 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -2754,22 +2754,45 @@ EXPORT_SYMBOL_GPL(call_rcu);
> #define KFREE_DRAIN_JIFFIES (HZ / 50)
> #define KFREE_N_BATCHES 2
>
> +/*
> + * This macro defines how many entries the "records" array
> + * will contain. It is based on the fact that the size of
> + * kfree_rcu_bulk_data structure becomes exactly one page.
> + */
> +#define KFREE_BULK_MAX_ENTR ((PAGE_SIZE / sizeof(void *)) - 2)
> +
> +/**
> + * struct kfree_rcu_bulk_data - single block to store kfree_rcu() pointers
> + * @nr_records: Number of active pointers in the array
> + * @records: Array of the kfree_rcu() pointers
> + * @next: Next bulk object in the block chain
> + */
> +struct kfree_rcu_bulk_data {
> + unsigned long nr_records;
> + void *records[KFREE_BULK_MAX_ENTR];
> + struct kfree_rcu_bulk_data *next;
> +};
> +
> /**
> * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
> * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
> * @head_free: List of kfree_rcu() objects waiting for a grace period
> + * @bhead_free: Bulk-List of kfree_rcu() objects waiting for a grace period
> * @krcp: Pointer to @kfree_rcu_cpu structure
> */
>
> struct kfree_rcu_cpu_work {
> struct rcu_work rcu_work;
> struct rcu_head *head_free;
> + struct kfree_rcu_bulk_data *bhead_free;
> struct kfree_rcu_cpu *krcp;
> };
>
> /**
> * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
> * @head: List of kfree_rcu() objects not yet waiting for a grace period
> + * @bhead: Bulk-List of kfree_rcu() objects not yet waiting for a grace period
> + * @bcached: Keeps at most one object for later reuse when build chain blocks
> * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
> * @lock: Synchronize access to this structure
> * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
> @@ -2783,6 +2806,8 @@ struct kfree_rcu_cpu_work {
> */
> struct kfree_rcu_cpu {
> struct rcu_head *head;
> + struct kfree_rcu_bulk_data *bhead;
> + struct kfree_rcu_bulk_data *bcached;
> struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
> spinlock_t lock;
> struct delayed_work monitor_work;
> @@ -2800,6 +2825,7 @@ static void kfree_rcu_work(struct work_struct *work)
> {
> unsigned long flags;
> struct rcu_head *head, *next;
> + struct kfree_rcu_bulk_data *bhead, *bnext;
> struct kfree_rcu_cpu *krcp;
> struct kfree_rcu_cpu_work *krwp;
>
> @@ -2809,22 +2835,39 @@ static void kfree_rcu_work(struct work_struct *work)
> spin_lock_irqsave(&krcp->lock, flags);
> head = krwp->head_free;
> krwp->head_free = NULL;
> + bhead = krwp->bhead_free;
> + krwp->bhead_free = NULL;
> spin_unlock_irqrestore(&krcp->lock, flags);
>
> - // List "head" is now private, so traverse locklessly.
> + /* List "bhead" is now private, so traverse locklessly. */
> + for (; bhead; bhead = bnext) {
> + bnext = bhead->next;
> +
> + rcu_lock_acquire(&rcu_callback_map);
> + kfree_bulk(bhead->nr_records, bhead->records);
> + rcu_lock_release(&rcu_callback_map);
> +
> + if (cmpxchg(&krcp->bcached, NULL, bhead))
> + free_page((unsigned long) bhead);
> +
> + cond_resched_tasks_rcu_qs();
> + }
> +
> + /*
> + * Emergency case only. It can happen under low memory
> + * condition when an allocation gets failed, so the "bulk"
> + * path can not be temporary maintained.
> + */
> for (; head; head = next) {
> unsigned long offset = (unsigned long)head->func;
>
> next = head->next;
> - // Potentially optimize with kfree_bulk in future.
> debug_rcu_head_unqueue(head);
> rcu_lock_acquire(&rcu_callback_map);
> trace_rcu_invoke_kfree_callback(rcu_state.name, head, offset);
>
> - if (!WARN_ON_ONCE(!__is_kfree_rcu_offset(offset))) {
> - /* Could be optimized with kfree_bulk() in future. */
> + if (!WARN_ON_ONCE(!__is_kfree_rcu_offset(offset)))
> kfree((void *)head - offset);
> - }
>
> rcu_lock_release(&rcu_callback_map);
> cond_resched_tasks_rcu_qs();
> @@ -2839,26 +2882,45 @@ static void kfree_rcu_work(struct work_struct *work)
> */
> static inline bool queue_kfree_rcu_work(struct kfree_rcu_cpu *krcp)
> {
> + struct kfree_rcu_cpu_work *krwp;
> + bool queued = false;
> int i;
> - struct kfree_rcu_cpu_work *krwp = NULL;
>
> lockdep_assert_held(&krcp->lock);
> - for (i = 0; i < KFREE_N_BATCHES; i++)
> - if (!krcp->krw_arr[i].head_free) {
> - krwp = &(krcp->krw_arr[i]);
> - break;
> - }
>
> - // If a previous RCU batch is in progress, we cannot immediately
> - // queue another one, so return false to tell caller to retry.
> - if (!krwp)
> - return false;
> + for (i = 0; i < KFREE_N_BATCHES; i++) {
> + krwp = &(krcp->krw_arr[i]);
>
> - krwp->head_free = krcp->head;
> - krcp->head = NULL;
> - INIT_RCU_WORK(&krwp->rcu_work, kfree_rcu_work);
> - queue_rcu_work(system_wq, &krwp->rcu_work);
> - return true;
> + /*
> + * Try to detach bhead or head and attach it over any
> + * available corresponding free channel. It can be that
> + * a previous RCU batch is in progress, it means that
> + * immediately to queue another one is not possible so
> + * return false to tell caller to retry.
> + */
> + if ((krcp->bhead && !krwp->bhead_free) ||
> + (krcp->head && !krwp->head_free)) {
> + if (!krwp->bhead_free) {
> + krwp->bhead_free = krcp->bhead;
> + krcp->bhead = NULL;
> + }
> +
> + if (!krwp->head_free) {
> + krwp->head_free = krcp->head;
> + krcp->head = NULL;
> + }
> +
> + /*
> + * The work can already be queued. If so, it means that
> + * within a short time, second, either head or bhead has
> + * been detached as well.
> + */
> + queue_rcu_work(system_wq, &krwp->rcu_work);
> + queued = true;
> + }
> + }
> +
> + return queued;
> }
>
> static inline void kfree_rcu_drain_unlock(struct kfree_rcu_cpu *krcp,
> @@ -2895,6 +2957,39 @@ static void kfree_rcu_monitor(struct work_struct *work)
> spin_unlock_irqrestore(&krcp->lock, flags);
> }
>
> +static inline bool
> +kfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, void *ptr)
> +{
> + struct kfree_rcu_bulk_data *bnode;
> +
> + if (unlikely(!krcp->initialized))
> + return false;
> +
> + lockdep_assert_held(&krcp->lock);
> +
> + /* Check if a new block is required. */
> + if (!krcp->bhead ||
> + krcp->bhead->nr_records == KFREE_BULK_MAX_ENTR) {
> + bnode = xchg(&krcp->bcached, NULL);
> + if (!bnode)
> + bnode = (struct kfree_rcu_bulk_data *)
> + __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
> +
> + /* No cache or an allocation got failed. */
> + if (unlikely(!bnode))
> + return false;
> +
> + /* Initialize the new block. */
> + bnode->nr_records = 0;
> + bnode->next = krcp->bhead;
> + krcp->bhead = bnode;
> + }
> +
> + /* Finally insert. */
> + krcp->bhead->records[krcp->bhead->nr_records++] = ptr;
> + return true;
> +}
> +
> /*
> * Queue a request for lazy invocation of kfree() after a grace period.
> *
> @@ -2926,9 +3021,17 @@ void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
> __func__, head);
> goto unlock_return;
> }
> - head->func = func;
> - head->next = krcp->head;
> - krcp->head = head;
> +
> + /*
> + * Under high memory pressure GFP_NOWAIT can fail,
> + * in that case the emergency path is maintained.
> + */
> + if (unlikely(!kfree_call_rcu_add_ptr_to_bulk(krcp,
> + (void *) head - (unsigned long) func))) {
> + head->func = func;
> + head->next = krcp->head;
> + krcp->head = head;
> + }
>
> // Set timer to drain after KFREE_DRAIN_JIFFIES.
> if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
> @@ -3834,8 +3937,11 @@ static void __init kfree_rcu_batch_init(void)
> struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
>
> spin_lock_init(&krcp->lock);
> - for (i = 0; i < KFREE_N_BATCHES; i++)
> + for (i = 0; i < KFREE_N_BATCHES; i++) {
> + INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
> krcp->krw_arr[i].krcp = krcp;
> + }
> +
> INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
> krcp->initialized = true;
> }
> --
> 2.20.1
>
next prev parent reply other threads:[~2020-01-16 1:14 UTC|newest]
Thread overview: 18+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-12-31 12:22 [PATCH 1/1] rcu/tree: support kfree_bulk() interface in kfree_rcu() Uladzislau Rezki (Sony)
2020-01-13 19:03 ` Paul E. McKenney
2020-01-14 16:49 ` Joel Fernandes
2020-01-15 13:14 ` Uladzislau Rezki
2020-01-15 22:53 ` Joel Fernandes
2020-01-17 17:52 ` Uladzislau Rezki
2020-01-17 18:57 ` Joel Fernandes
2020-01-17 21:37 ` Paul E. McKenney
2020-01-17 21:59 ` Joel Fernandes
2020-01-19 13:03 ` Uladzislau Rezki
2020-01-16 1:14 ` Joel Fernandes [this message]
2020-01-16 2:41 ` Paul E. McKenney
2020-01-16 17:27 ` Uladzislau Rezki
2020-01-16 17:44 ` Paul E. McKenney
2020-01-16 17:24 ` Uladzislau Rezki
-- strict thread matches above, loose matches on Subject: below --
2019-12-20 12:56 Uladzislau Rezki (Sony)
2019-12-21 23:21 ` Joel Fernandes
2019-12-24 18:49 ` Uladzislau Rezki
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20200116011410.GC246464@google.com \
--to=joel@joelfernandes.org \
--cc=linux-kernel@vger.kernel.org \
--cc=oleksiy.avramchenko@sonymobile.com \
--cc=paulmck@kernel.org \
--cc=rcu@vger.kernel.org \
--cc=rostedt@goodmis.org \
--cc=urezki@gmail.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).